Meet Jev, the AI that claims it can’t hallucinate

Self-proclaimed ChatGPT “co-creator” Diogo Almeida is running a viral marketing campaign for a new AI model, Jev, based on his work on the OpenAI chatbot’s core function, Reinforcement Learning from Human Feedback (RLHF).

The problem is, he’s admitting that his own invention doesn’t work well.

Almeida’s new startup, TypeSafe, emerged this month with a $40 million seed round, and a slick pitch that claimed he’d solved a problem that created during his time working on ChatGPT: AI hallucinations.

The pitch for its Jev model, however, is much less impressive and raises a few questions.

First of all, the claim that he “co-created ChatGPT” isn’t strictly true.

In fact, Almeida is the fourth of 20 listed authors on the 2022 InstructGPT paper that applied RLHF to language models. This was one research paper that ChatGPT engineers used while building the chatbot.

Relegating his role further, a footnote in that PDF marks Almeida as one of nine “primary” authors rather than a team lead.

Moreover, RLHF technology predates that paper by nearly five years and its foundational method came from six other researchers.

Jev, an AI built on a résumé failure

Almeida is now attempting to launch and raise funding for Jev by admitting that his own creation, RLHF, “is not useful.”

He calls ChatGPT “unreliable without human supervision,” and he’s blamed RLHF for “mode dropping, overconfidence, and an overall lack of reliability.”

Almeida is also taking the opportunity to claim that his new AI model “can’t hallucinate.”

His claim about anti-hallucination technology, however, falls apart after even a momentary consideration.

Specifically, Jev’s supposed protection against hallucinations, according to its own documentation, is simply its option-choosing rather than sentence-writing design with a hard-coded cap of 255 options, not any particularly newsworthy code or algorithmic breakthrough.

Jev literally and quite simplistically presents itself as a simple decision selector among a limited set of options or pre-determined choices such as determining a likelihood percentage.

Because it doesn’t deliver text through a traditional chatbot, it literally cannot hallucinate because hallucinations require phrases that make sentences or factual claims.

Jev can only choose from predefined outputs and doesn’t deliver sentences that it or humans can misinterpret. Almeida therefore boasts that hallucinations in Jev are impossible by design.

Jev is days old and still in early access. It prices input tokens at $42 per billion, with output tokens momentarily free amid subsidies. Almeida claims speed, cost, and quality advantages over competitors.